Bilkent University
Department of Computer Engineering


Diversity Measures in Online Ensemble Learners


Nurettin Aykut Güven
MS Student
Computer Engineering Department
Bilkent University

Ensemble learners is a committee of classifiers gathered to overcome adversaries in single learners, which is commonly used in data stream applications. Diversity among ensemble members is regarded as one of the key points in an ensemble’s structure and accuracy, as Condorcet’s theorem implies. However, diversity measurements are not straightforward as there is no commonly accepted formal definition. In this study, we propose a novel straightforward diversity measure for ensemble classifiers based on a previous study in our research group that offers a geometrical framework for online ensemble learners. The divRank measure, that stands for “diversity rank,” is the ratio of the rank of the coefficient matrix of the score vectors of an ensemble to the ensemble size. We measured how the divRank of an ensemble changes with respect to different combinations of ensemble size and the number of class labels. Different diversity measures and ensemble topologies were taken as baselines. Our first results indicated that as the number of identical or correlated classifiers increases, divRank value decreases, and that linear dependence between the score vectors of ensemble members can be correlated to the definition of diversity.


DATE: 05 March, 2018, Monday, CS590 & CS690 presentations begin at @ 15:40